diff --git a/python/sglang/srt/batch_overlap/two_batch_overlap.py b/python/sglang/srt/batch_overlap/two_batch_overlap.py
index 14da7ad9e..9d5a96b85 100644
--- a/python/sglang/srt/batch_overlap/two_batch_overlap.py
+++ b/python/sglang/srt/batch_overlap/two_batch_overlap.py
@@ -701,12 +701,20 @@ class TboForwardBatchPreparer:
             "spec_algorithm",
             "capture_hidden_mode",
             "padded_static_len",
-            "mrope_positions",  # only used by qwen2-vl, thus not care
             "split_index",  # for split prefill
             "orig_seq_lens",  # only used by qwen-1m, thus not care
             "return_pooled_hidden_states",
         ]:
             output_dict[key] = getattr(batch, key)
+
+        mrope_positions = getattr(batch, "mrope_positions")
+        if mrope_positions is not None:
+            output_dict["mrope_positions"] = mrope_positions[
+                :, start_token_index:end_token_index
+            ]
+        else:
+            output_dict["mrope_positions"] = None
+
         if not batch.forward_mode.is_target_verify():
             assert (
                 _compute_extend_num_tokens(batch.input_ids, batch.forward_mode)
diff --git a/python/sglang/srt/layers/attention/tbo_backend.py b/python/sglang/srt/layers/attention/tbo_backend.py
index 494d82d80..2ae120686 100644
--- a/python/sglang/srt/layers/attention/tbo_backend.py
+++ b/python/sglang/srt/layers/attention/tbo_backend.py
@@ -179,6 +179,9 @@ class TboAttnBackend(AttentionBackend):
             assert ans == child.get_cuda_graph_seq_len_fill_value()
         return ans
 
+    def forward(self, *args, **kwargs):
+        return self.primary.forward(*args, **kwargs)
+
     def forward_extend(self, *args, **kwargs):
         return self.primary.forward_extend(*args, **kwargs)